The Effect of Elections and Prime Ministers on Discussion in the Australian Federal Parliament (1901-2018) - Rohan Alexander

 
CONTINUE READING
The Effect of Elections and Prime Ministers on Discussion in the Australian Federal Parliament (1901-2018) - Rohan Alexander
The Effect of Elections and Prime
Ministers on Discussion in the Australian
Federal Parliament (1901–2018)
Monica Alexander, University of Toronto
Rohan Alexander, Australian National University

Annual Conference of The Political Methodology Specialist Group
University of Warwick
11 January 2019

Thank you to the ANU RSE for the funding to attend this conference
The Effect of Elections and Prime Ministers on Discussion in the Australian Federal Parliament (1901-2018) - Rohan Alexander
Introduction

Data

Topics

Model

Results

Conclusion

               !2
Motivation
Elections and changes in prime minister
occur regularly and are major events,
but the broad conditions often remain
the same.
Motivation

 “When the government changes,
     the country changes”
 Paul Keating on the danger of himself being replaced
           by John Howard (as quoted by Paul Kelly).

“There's no such thing as changing
 the government without changing
 the circumstances of the country”
 John Howard on the danger of himself being replaced
     by Kevin Rudd (as quoted by Peter van Onselen).

                                                        Paul Keating by Robert Lyall Hannaford.
                                                                                              !4
Motivation

In Australia the prime minister is
decided by the party in power and
can be replaced without an election.
Similarly, a re-elected government
can either ‘reboot’ or focus on
stability after being returned. But the
party in power, much of the cabinet,
and general economic conditions,
remained unchanged despite either
of these significant events.

                                          Julia Gillard by Vincent Fantauzzo.
                                                                            !5
Questions
Do elections affect the topics discussed
in the Australian Parliament House? For
instance, does John Howard change
focus after each election win?

Similarly, does a changed prime minister
affect the topics discussed in the
Australian Parliament House? For
instance, is Paul Keating much different
to Bob Hawke?

Has this changed over time? For
instance, is John Howard’s 1996–2007
different to Robert Menzies’ 1949–1966?
                                           John Howard by Jiawei Shen.
                                                                     !6
Approach
We first summarise what was said in
the Australian Parliament House from
Federation in 1901 though to 2018
using a correlated topic model.

Then analyse the changes in those
topics using a Bayesian Dirichlet
model with government and election
levels effects.

We are looking neighbouring prime
ministerial or election periods where
the ‘mix’ of topics is different.
                                        Bob Hawke by William (Bill) Leak.
                                                                        !7
Findings
Prime minister
Changes in prime minister tend be
associated with topic changes even when
the party in power does not change.

Elections
Elections where the party in power also
changes are associated with topic
changes.

Timing
Since the 1990s re-elections begin to be
associated with a significant change in
topics.
                                           Malcolm Fraser by Sir Ivor Henry Hele.
                                                                                !8
Contributions
Data
We bring an essentially-complete new corpus
of who said what in the Australian Federal
Parliament on a daily basis.

Methods
We introduce an alternative to the STM
approach that: 1) allows more-complicated
auto-correlated functional forms; 2) implements
pooling across groups of similar days; and 3)
identifies outlying topic distributions without
the need to pre-specify the event of interest.

Australian political knowledge
We show one way in which Australian politics
has changed over time.
                                                  Gough Whitlam by Clifton Ernest Pugh.
                                                                                      !9
Introduction

Data

Topics

Model

Results

Conclusion

               !10
Data
Hansard PDFs are available since
Federation (1901). XML available, but
incomplete. No turnkey Hansard
corpus for Australian researchers, yet.

Creating a corpus required a large
PDF-parsing and data-cleaning
exercise. We end up with 7,934 days in
House of Representations and 6,746
days in Senate, across 118 years.

Our CSV corpus (c.4GB) is available
for other researchers.                    First page of Hansard for 13 July 1906 in
                                                    the House of Representations.

                                                                                  !11
Introduction

Data

Topics

Model

Results

Conclusion

               !12
Vignette
 “The Opposition wages policy is a joke and would be washed to one side in the real economy, if it
  were ever put into place. Yet Opposition members talk about the Opposition's credentials… the
                        Opposition crowd could not raffle a duck in a pub.”
                                                                                     Paul Keating, 16 September 1986.

Word counts
      date             the             opposition        wages              policy                     …
  1986-09-16            3                  4               1                  1                        …
       …                …                 …               …                  …                         …

Topic proportions
       date         Topic 1: economy, wages, policy, …    Topic 2: opposition, opposite, joke, …              …
   1986-09-16                      0.6                                      0.4                               …
       …                           …                                       …                                  …

                                                                                                                  !13
Topic model output
We use a Correlated Topic Model          Chamber Date      Topic Proportion
(CTM) and specify 80 topics. CTM         HoR     1901-05-09 1       1.4889e-04
output is a proportion for each of the
                                         HoR     1901-05-10 1       3.2654e-04
80 topics for each day.
                                         HoR     1901-05-21 1       1.8766e-04

Topics are defined by collections of     HoR     1901-05-22 1       1.1172e-04
words, e.g. Topic 4: defence, forces,    HoR     1901-05-23 1       2.4848e-03
personnel, army, military,
                                         HoR     1901-05-29 1       3.7861e-03
defence_force, equipment, base,
aircraft, air…; and Topic 74: budget,    HoR     1901-05-30 1       2.4947e-03
tax, billion, million, per_cent,         HoR     1901-05-31 1       2.8733e-05
business, economy, support, jobs,        …       …         …        …
governments…                                              Output from a topic model.

                                                                                   !14
Introduction

Data

Topics

Model

Results

Conclusion

               !15
Model
The topic, p, proportion on some day, d, (e.g. the 0.6 in the example) is θc,d,1:P :

                          θc,d,1:P ∼ Dirichlet(μc,s[d],1:P)

                        log μc,s,p = αg,p + βe,p ⋅ s + δc,s,p

The Dirichlet distribution is useful when we have proportions and more than
two categories. Our model considers prime minister and election levels
effects and sitting period random effects.

                                                                                 !16
Prime minister levels effects

                       log μc,s,p = αg,p+βe,p ⋅ s + δc,s,p

 The term for the prime minister assumes there is some underlying mean
 effect of each prime minister on the topic distribution. We place
 uninformative priors on each of these parameters:

                              αg,p ∼ N(0,100) .

                                                                         !17
Election levels effects

                        log μc,s,p = αg,p+βe,p ⋅ s+δc,s,p

 The election term assumes there is some effect of an election on the
 topic distribution. This effect decays the further away from the election
 some sitting period is. Again, uninformative priors:

                              βe,p ∼ N(0,100) .

                                                                            !18
Sitting period random effects
                       log μc,s,p = αg,p + βe,p ⋅ s+δc,s,p

 The sitting-period-specific random effect allows the topic distributions in
 some sitting periods to be different than would be expected based on
 the prime minister and election effects. This allows us to identify large
 deviations away from the expected distribution, thus helping to identify
 the effect of other, non-prime-minister and non-election events. Also, this
 set-up partially pools effects across sitting periods.

                                               2
                               δc,s,p =   N(0,σg,p)

                                 2
                                σg,p   ∼ U(0,3)

                                                                              !19
Introduction

Data

Topics

Model

Results

Conclusion

               !20
Significant elections and PMs
                                         Elections

                   ● ●               ●    ●● ●          ● ●    ● ●● ● ●       ● ● ● ● ● ● ●

                                           PMs

          ●   ●●   ●     ●   ●   ●   ●            ● ● ●● ●       ●        ●   ●      ● ● ●

   1900                          1950                                         2000
                                                  Each election/PM is a grey dot. Those that were
                                           significantly different to their predecessor are in black.
                                                                                                       !21
Introduction

Data

Topics

Model

Results

Conclusion

               !22
Summary
1. Collect, parse, and clean Australian Hansard PDFs to construct a corpus
   with around 15,000 days spread across 118 years.

2. Summarise the text using a topic model, which provides a quantitative
   sense of what was talked about each day.

3. Analyse the topic model output using a Bayesian Dirichlet model to look for
   changes in what is talked about based on who is the prime minister and
   which election period we are in.

4. Find that a different prime minister tends to bring a change in the mix of
   what is talked about, but that elections have mostly only been associated
   with a change since the 1990s.

                                                                               !23
Weaknesses
Data
Even after cleaning the dataset remains
imperfect and is more fit-for-purpose than
of broad applicability.

Topics
The number of topics needs to be tractable
for the model, but more topics would be
better based on usual topic-model
diagnostics.

Model
We consider a two-stage process, but do
not appropriately propagate the uncertainty
of the topic distribution estimation stage.
                                              Photo of the Harold Holt Memorial Pool by Kbpool2012,
                                              from Wikimedia Commons.
                                                                                                 !24
Research agenda
“But where are you really from? The Changing Effect of State and Party on
Senators’ Discussion in the Australian Parliament House (1901—2018)” with
Patrick Leslie. We analyse how what is said in the Senate is affected by the
state the senator represents and how this has changed over time. Directly
considers words:
               ci,t ∼ multinomial(ni,t, pi,t)
             pi,t = αi + β1,t × chamberi + β2,t × party+β3,t × state

“Your house or mine? The Changing Focus of the Australian Colonial
Parliaments (1880—1920)” with Tim Hatton. We examine the changing focus of
the colonial (later state) parliaments before and after Federation.

                                                                              !25
“The Effect of Elections and Prime Ministers on Discussion in the Australian
Federal Parliament (1901–2018)”
Monica Alexander and Rohan Alexander

Email: rohanalexander@anu.edu.au.

Twitter: @rohanalexander.

Paper available at: rohanalexander.com/academic.

Data freely available for download and use, but maybe contact me if you need to know where the bodies
are buried.

Acknowledgements: Thank you to Chris Cochrane, Dan Simpson, Jill Sheppard, John McAndrews, John
Tang, Leslie Root, Martine Mariotti, Matt Jacob, Matthew Kerby, Myles Clark, Ruth Howlett, Tianyi Wang,
Tim Hatton, and Zach Ward for their invaluable contributions; and to the UC Berkeley Demography
Department for the use of their computing resources. We are grateful for the many excellent comments
that we received from seminar participants at the ANU SPIR, the ANU RSE, the Australian Parliamentary
Library, the Max Planck Institute for Demographic Research, and the U of T Political Behavior Group.

Slides theme based on Nathan Lane, see https://slides.com/nathanlane/kdi#/.

                                                                                                      !26
You can also read